Controlled studies of customer support agents, professional writers, and consultants show real, measurable gains from AI. They also show where it falls short.

There is no shortage of bold claims about what AI can do for a business. The more useful question is narrower: when researchers actually measure people doing real work with and without AI, what changes?

A handful of well-designed studies now answer that question. They were run by economists at MIT, Stanford, Harvard, and the Federal Reserve, and their results are consistent enough to be worth taking seriously.

Customer support: 14% more issues resolved per hour

In one of the largest real-world studies to date, researchers followed 5,179 customer support agents at a Fortune 500 software company as an AI assistant was rolled out. Agents with access to the tool resolved 14% more issues per hour on average, according to the paper by Erik Brynjolfsson, Danielle Li, and Lindsey Raymond, later published in the Quarterly Journal of Economics.

The gains were not evenly spread. Newer and less-skilled agents improved by 34%, while the most experienced agents saw little change. The researchers concluded that the AI was effectively passing along the habits of top performers to everyone else. They also found that customer sentiment improved and agents were more likely to stay with the company.

Professional writing: 40% faster, 18% better

A preregistered experiment published in Science gave 453 college-educated professionals (marketers, consultants, data analysts, HR staff) realistic writing tasks such as press releases, short reports, and workplace emails. Half were given access to ChatGPT.

Those with AI finished about 40% faster, and independent graders rated their work 18% higher in quality. As with the support agents, the weakest writers improved the most, narrowing the gap between workers.

Consulting work: big gains inside the "frontier," losses outside it

Researchers from Harvard Business School, Wharton, MIT Sloan, and Boston Consulting Group ran an experiment with 758 BCG consultants. On tasks well suited to AI, consultants using GPT-4 completed 12.2% more tasks, worked 25.1% faster, and produced results rated more than 40% higher in quality.

But on a task deliberately chosen to fall outside what the AI did well, consultants using it were 19 percentage points less likely to reach the correct answer than those working without it. The authors called this the "jagged technological frontier": AI is very good at some tasks and confidently wrong on others that look similar, and the line between them is not always obvious.

Across the whole workforce: real, but more modest

Controlled experiments show what is possible on well-chosen tasks. A nationally representative survey by the Federal Reserve Bank of St. Louis shows what is happening on average. Workers who used generative AI reported saving about 5.4% of their work hours, roughly 2.2 hours in a 40-hour week. About one in five users saved four or more hours a week.

Spread across all workers, including those who do not use AI at all, the savings came to 1.4% of total hours. The authors estimated that this corresponds to about a 1.1% increase in aggregate productivity. The gap between those two numbers is the opportunity: the benefit depends heavily on whether AI is actually built into the work people do every day.

What this means for a business

Taken together, the evidence points to a few practical conclusions:

  • The biggest gains come on well-defined, repetitive, text-heavy work: answering common questions, drafting routine documents, summarizing information.
  • AI tends to help newer and less experienced staff the most, which makes it a useful tool for onboarding and consistency.
  • It is not reliable everywhere. Work that falls outside its strengths needs human judgment, and outputs should be reviewed rather than trusted blindly.
  • Results depend on adoption. A tool used daily inside a real workflow delivers far more than one that people open occasionally.